The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads from a sensitive local path ~/Library/Application Support/MacWhisper/Database/main.sqlite. This database contains private meeting transcriptions, speaker labels, and session metadata.
[COMMAND_EXECUTION]: The skill invokes the system sqlite3 command-line utility using child_process.spawnSync to interact with the local transcription database.
[COMMAND_EXECUTION]: Database queries in lib/db.js (such as searchSessions and getTranscriptLines) are constructed using string interpolation of user-supplied inputs from command-line arguments. While basic single-quote escaping is applied in the search function, this pattern is less secure than using parameterized queries and could lead to SQL injection vulnerabilities.
[PROMPT_INJECTION]: The skill ingests and processes untrusted transcription data (Indirect Prompt Injection surface).
Ingestion points: lib/db.js (queries the main.sqlite database).
Boundary markers: None; the transcript is returned as a raw string without explicit delimiters or warnings to the agent to ignore embedded instructions.
Capability inventory: spawnSync (runs the sqlite3 CLI) and fs.writeFileSync (manages a local state file).
Sanitization: The skill performs "hallucination filtering" to remove non-Latin scripts and screen-reading artifacts, but it does not sanitize the text for potential malicious instructions intended to influence the agent's behavior.

macwhisper